fix(agent): prefer native vision over auxiliary fallback in auto mode (#29135)

[teknium1]

Summary

image_input_mode: auto now prefers native vision when the main model supports it; an explicit auxiliary.vision config acts as a fallback for text-only main models instead of unconditionally forcing the lossy text pipeline. Fixes #29135 and the bug half of #44299 (also reported in #53400 / #48964).

Root cause: decide_image_input_mode() checked _explicit_aux_vision_override(cfg) before _lookup_supports_vision(), so any user with an auxiliary.vision block configured lost native pixels even on a fully vision-capable main model.

Changes

• agent/image_routing.py: reorder auto-mode decision — native vision capability wins; aux override is the fallback for non-vision models. Docstrings updated to match — cherry-picked from fix(agent): prefer native vision over auxiliary fallback in auto mode (#29135) #32519 (@Jiahui-Gu)
• tests/agent/test_image_routing.py: updated decision-table coverage for both orderings

Validation

Test file	Result
tests/agent/test_image_routing.py	pass
tests/agent/test_custom_providers_vision.py	pass

Closes #29135, #44299, #53400, #48964. Salvages #32519 (canonical earliest PR; #44320, #44305, #44450 were duplicates already closed against it) with contributor authorship preserved.

Infographic

[alt-glitch]

This was generated by AI during triage.

Core-team salvage (@teknium1) of the earliest canonical PR #32519 (@Jiahui-Gu), authorship preserved. Closes the bug/feature cluster #29135, #44299, #53400, #48964. Distinct from the additive native_first opt-in mode PR #57297 (new mode vs. reordering auto). Related, not duplicate — this is the canonical fix; the salvaged/competing items can be closed in its favor by a maintainer.